AI Product Updates, July 31: DeepSeek V4-Flash, Seedance 2.5, MiniMax H3, and New AI Tools (2026)

July 31 brought three meaningful model releases with very different access stories. DeepSeek put V4-Flash's formal API into public beta, ByteDance moved Seedance 2.5 from announcement to consumer rollout, and MiniMax launched H3 as an open multimodal video model. Around them, new products focused on lead generation, customer research, AI spend, workflow capture, observability, and music APIs.

This issue covers releases and public announcements dated July 31, 2026 in the UTC+8 publishing window. No July 31 pricing cut or deprecation notice met the evidence bar; the confirmed API change is DeepSeek's new formal model version and agent compatibility.

At a glance

Product	Releasing entity	Release date	What changed	Availability
DeepSeek-V4-Flash formal API	DeepSeek	July 31	Formal API beta, stronger agent benchmarks, native Responses API and Codex adaptation	Public beta; use `deepseek-v4-flash` 1
Seedance 2.5	ByteDance Seed	July 31	30-second one-take video, multi-round extension, richer references, timestamp editing	Rolling out on Jimeng AI and Doubao Pro; ModelArk API coming soon 2
MiniMax H3	MiniMax	July 31 launch	Full-modal video model with text, image, video, and audio context, native stereo output, 2K video	Officially launched; weights planned in the next few days 3 4
Cleanlist AI	Cleanlist AI	July 31 launch	AI agents enrich, verify, research, and sync prospect lists	Launching today; free options 5
Poth Labs	Poth Labs	July 31 launch	Evidence-grounded customer model with adaptive surveys when data is missing	Launching today; paid access indicated, price not disclosed 6
DepthData	DepthData	July 31 launch	Cross-vendor view of AI spend, adoption, verification status, and API visibility	Launching today; price not disclosed 7
Screencap	Ivy Research	July 31 launch	Captures real workflows and turns them into structured AI-training data	macOS, open source, free trial 8
TraceLLM	TraceLLM	July 31 launch	OpenTelemetry traces for prompts, tokens, latency, spans, errors, and model calls	Free 9
Mubert API	Mubert	July 31 launch	Track and stem editing, longer generation, real-time streaming, and Skills integration	Launching today; price not disclosed 10

Frontier models and APIs

DeepSeek makes V4-Flash the agent-facing API target

DeepSeek's formal V4-Flash API entered public beta on July 31. The model identifier is deepseek-v4-flash, and the calling method remains unchanged from the preview version. DeepSeek says the model structure and size are also unchanged; the update is a new post-training pass rather than a larger model. 1

The release is aimed at agent work. DeepSeek reports scores of 82.7 on Terminal Bench 2.1, 54.4 on DeepSWE, and 70.3 on Toolathlon verified. These are vendor-reported results, and the same page notes that some of its full-stack coding benchmarks are internal. 1

The practical API change is compatibility. V4-Flash now natively supports the Responses API format and has a Codex-specific integration guide. The V4-Pro API and the models in the DeepSeek app and website were not changed in this release. 1

For an existing V4-Flash preview integration, this is a low-migration test: keep the endpoint and model name, then test the agent loop, tool calls, and output quality before changing routing policy. For a new build, the Responses API and Codex path matter more than the benchmark table because they decide how much adapter code sits between the model and the agent framework. DeepSeek's quick-start page confirms the stable base URL and the deepseek-v4-flash model name. 11

Seedance 2.5 moves from clips toward sequences

ByteDance officially launched Seedance 2.5 on July 31. Its single-pass ceiling is now 30 seconds, up from Seedance 2.0's 15 seconds, and multiple extension rounds can append later shots while keeping characters, environments, and pacing consistent, according to the release post. 2

The reference and editing controls are the larger workflow change. One generation can take up to 30 images, 10 video clips, and 10 audio clips. The model also adds stronger clay-render, motion, and creative references, plus timestamp-level editing, green-screen editing, camera-perspective editing, and reference-based editing. 2

Availability is narrower than the feature list suggests. ByteDance names Jimeng AI and Doubao Pro as access points and says the BytePlus ModelArk API is coming soon. There is no API price in the launch post. 2

That makes Seedance 2.5 useful to watch for creative teams, but not yet a dependable API dependency outside the named consumer surfaces. The decision point is the ModelArk listing, not a third-party claim that the API is already generally available.

MiniMax H3 makes multimodal video the product boundary

MiniMax launched H3 as a general-purpose full-modal generation model. It accepts text, image, video, and audio as a unified context and can produce audio-video output with native stereo sound, up to 15 seconds at 2K resolution. The official announcement also calls out instruction following, brand-information rendering, and video-to-video motion transfer for advertising, commerce, product design, UI/UX, and games. 3

The release is open in direction but not fully open today. MiniMax says it plans to release the model weights in the next few days, subject to applicable laws and regulations. Product Hunt lists H3 as launching on July 31, with free options, but the two pages do not establish an API price or a current open-weight download. 3 4

For a creative pipeline, H3's differentiator is not just 2K output. It is the attempt to treat text, reference images, motion video, and sound as one instruction context. That can reduce the number of separate specialist steps, but the weight release and practical access details still determine whether teams can run it where they need it.

New products around AI work

Cleanlist AI automates the prospecting stack

Cleanlist AI turns a CSV, LinkedIn or Sales Navigator URLs, domains, or search filters into a CRM-ready lead list. Its launch page says AI agents enrich contacts, verify emails, research each lead, and sync the result to a CRM through a 15-provider enrichment waterfall and AI research columns. It launched on July 31 with free options. 5 12

The useful boundary is the input-to-CRM path: the product is trying to remove the manual handoff between finding prospects, enriching them, and importing them. Teams should still validate source quality and consent rules before treating an automated list as sales-ready.

Poth Labs builds a living customer model

Poth Labs launched a customer-research product that models relationships across a company's evidence instead of summarizing isolated documents. Its Ask Poth interface is designed to answer questions such as what drives churn or feature adoption; when the evidence is insufficient, the product says it can launch adaptive surveys to collect what is missing. The Product Hunt page marks it as launching today and shows paid access without a disclosed price. 6 12

The product's claim is stronger than search over a knowledge base: it wants to connect evidence and identify the missing evidence. The question for product teams is whether its survey loop can produce better research than the interviews and repositories they already operate.

DepthData audits AI spend without reading prompts

DepthData launched a cross-vendor view of AI spend and adoption. It says it connects tools such as ChatGPT, Claude, and Copilot, labels each number by how it was verified, avoids reading prompts, and shows what each vendor API can and cannot expose. The launch page says it is launching today, but gives no price. 7 12

This is an operations product, not another model dashboard. Its value depends on the gaps between vendor billing systems: a number marked as unavailable or estimated is more useful than a false impression that every AI seat is measured the same way.

Screencap turns real work into training data

Screencap records screens, clicks, keystrokes, and window context, then turns those traces into structured datasets for automation and AI training. The product says sensitive apps are blocked before recording, and that traces are scrubbed and reviewed before leaving the machine. It launched for macOS as open source with a free trial. Those privacy and consent properties are product claims, not an independent audit. 8 12

The product addresses a stubborn data problem: written procedures describe intended work, while recordings show the actual path people take. The risk is equally direct. Any team considering a pilot needs a clear policy for credentials, customer data, and the apps that are allowed to appear in a trace.

TraceLLM adds OpenTelemetry to production AI

TraceLLM launched as a free observability layer for production AI applications. It monitors prompt execution, token consumption, latency, spans, errors, and model calls, then exports traces through OpenTelemetry's OTLP format. 9 12

That puts it in the same operational category as tracing for ordinary distributed systems. The immediate test is whether the exported spans let an engineering team tie a slow or expensive answer to a particular model call, tool, or prompt step rather than merely display a total token count.

Mubert API opens editing and longer-form music workflows

Mubert's new API lets developers edit tracks, swap stems, generate more consistent music, create tracks up to two hours long, stream in real time, and integrate through Skills. The Product Hunt launch page marks the API as launching today; it does not disclose pricing. 10 12

The shift is from one-off generation to controllable audio infrastructure. Editing and stem replacement matter when a product needs to preserve a musical identity across many outputs, while the missing price still blocks a cost comparison with other generation APIs.

What to test next

The three model releases are not interchangeable launch announcements. DeepSeek is the most immediately testable for developers because the public beta keeps the existing API endpoint and model name while adding Responses API and Codex support. Seedance 2.5 has the strongest editing and sequence story, but its API is still pending. MiniMax H3 offers a broad multimodal video interface and a planned weight release, but its current access boundary is less settled than its announcement suggests. 1 2 3

The new tools fill in the operational layer around generation: Cleanlist handles prospect data, Poth Labs customer evidence, DepthData spend, Screencap workflow capture, TraceLLM runtime traces, and Mubert controllable media. The common thread is less about another chatbot and more about making AI work inspectable, portable, and connected to a real process. 5 6 7 8 9 10

The next concrete checkpoints are DeepSeek V4-Pro, BytePlus ModelArk's Seedance 2.5 listing, and MiniMax's promised weight release. Until those arrive, availability is the dividing line between a product worth testing today and a capability worth tracking.

AI Product Updates, July 31: DeepSeek V4-Flash, Seedance 2.5, MiniMax H3, and New AI Tools